Almost half of us use AI for info and ideas. Easy to see why. Social media is rot. Google is just a portal to Reddit now. We’re hungry for reliability. Plus, aren’t bots helpful? I asked one if it knew it wasted power. Got a vegan cheese recipe 30 minutes later.
Did not make the cheese. Found a human-made recipe it likely stole. That’s how they work. They repack knowledge into tailored slop. Fine for food. Not for truth. I am a fact-checker for WIRED. Stakes are higher.
People pity me lately. They think I’m obsolete. Foolish. I think very little human knowledge actually lives on the internet. AI is worse than people think. Wronger.
Colin Dickey says Tom Wolfe viewed us fact-checkers as a cabal of henpeckers. Fair. My boss is a man. We are annoying. That’s the job.
Old-School
WIRED does this the hard way. Line by line. Primary sources. We call people. We wait on hold. We argue with ethics and lawyers. It is a peer review on speed.
AI hasn’t killed this. Yet. It is trying for post hoc checks. Like Snopes but automated. In the UK, a group called Full Fact built tools for this. They scan posts and podcasts. They flag claims. Then humans investigate. Mark Frankel of Full Fact gets it. You need a human.
“You definitely need a human being.”
How Wrong Is Wrong?
AI fails. Often. How often? Hard to say. 17,00 papers on arXiv since 2018. Mostly about reliability.
At the desk, we check b-matter. Stats. Dates. Quotes. I use AI Overviews. I hate them. Wrong about a third of the time. Maybe worse.
A Tow Center study in March 2025 says over 60% of AI search results are inaccurate. BBC says 45%. Let’s be blunt. It’s half wrong.
Which model? Musk thinks Grok is best. Research disagrees. Claude topped RealFactBench last year at 73% accuracy. Grok wasn’t even in it. SimpleQA, run by OpenAI in Oct 2024? No model beat 50%. Google updated it recently. Gemini 2.5 won with 55.6%.
ChatGPT? It told me models are 90-96% accurate. Then linked a paper on sleep medicine. It said hallucinations happen 1-2% of the time. The source didn’t exist.
Smarter doesn’t mean truer. Sometimes it’s the opposite. A 2025 AI report found 60% of researchers doubt factuality will be solved soon. The bots try too hard to please. They invent instead of admit defeat.
The Test
I have the hiring test I passed years ago. Fake story. Robocalling kingpin. Three bonus questions. I gave it to ChatGPT. Claude. Gemini. Grok. Free versions.
Grok was aggressive. “Yes.” It hated truth. Wanted infinite data. It noted fact-checking is historically women’s work. Interesting. Useless.
Claude and Gemini were decent. Reasonable plans. Legal flags. Gemini wanted “Paper Trails” for “People Trails”. Cringe.
ChatGPT was eager. Insecure. Buzzwords. It wanted to diagram sentences. Then it made up a paragraph from the story. None of them actually checked facts. They promised plans. Then stopped.
“I don’t think it’s an option… to sit AI out,” says Angie Holan from Poynter. She prefers engagement. Learn the tools. See their breaks.
I agree. I feel sharper now.
The Analog Advantage
The fun starts when Google fails. A sign at a border. Kelp growth rates. A Burger King in 1979 LA. These things are ghosts on the net.
A bot won’t sit on the phone for an hour while a widow cries. It won’t care about the grief. It can’t feel passive hostility in “Thanks for your email.” It can’t see the beef between sources blurring facts.
Jack Bialik wrote that we think old tech is new. Assembly lines. Cataract surgery. Our digital storage rots. Microchips last five years. Stone lasts millennia.
Ada Palmer says we know less than 1% of history from 500 years ago. And two-thirds is wrong. Generations passed it down. Bits were lost. Now we trust servers? Servers die.
Humans make mistakes. I know this. Holan said not using chatbots isn’t safety. At least… I think she said that. 33% to 90% sure.
I checked the recorder at the end of our interview.
It was off.
Your Turn
Comments are below. Email us at [email protected] if you hate the vibe. Or love it.
